The future of decentralized finance (DeFi) has gone beyond just smart contracts with the mass adoption of artificial intelligence (AI). There is now a growing need to embrace the potential of combining DeFi mechanics with AI autonomy.
They are capable of trading, liquidity management, risk audits, and decentralized autonomous organization (DAO) governance, all in real-time and without a human having to press a button. Therefore, it is crucial to understand how to deploy AI agents using integrated large language models (LLMs), particularly when building in Web3 or DeFi.
An on-chain AI agent is an autonomous program that utilizes an LLM for critical thinking, planning, and decision-making. The program outputs are then executed via smart contracts on the blockchain. Unlike traditional bots that follow pre-defined sets of rules, the on-chain AI agent can interpret natural language commands, analyze market data, and interact with the blockchain.
An on-chain AI agent comprises three components. First, there is the intelligence layer that utilizes the LLM (whether it is GPT-4o, Claude Sonnet, or an open-source alternative such as Mistral). The execution component, which employs a smart contract or a wallet system. Finally, a middleware that acts as a tool framework that connects the LLM to the blockchain.
Before writing any code, you need to make three decisions.
1. Select your LLM
For most DeFi applications, an API from OpenAI or Anthropic is one seamless way to start. For more privacy-focused or decentralized use, an open-source model hosted on a decentralized compute network (such as 0G Labs or Internet Computer Protocol), which has demonstrated full on-chain model hosting, may be more suitable.
2. Choose a blockchain
A blockchain of choice is one with a minimal transaction fee and Coinbase CDP SDK integrations. A widely accepted example is Base Chain, which uses the Ethereum blockchain’s OP Stack. Another option is the Ethereum mainnet blockchain; however, the gas fee is relatively higher. For multi-chain operations, you will need cross-chain bridges built into your agent logic.
3. Pick a framework
This is the infrastructure that handles memory, tool routing, and multi-step reasoning chains. The top open-source frameworks for building agent workflows are LangChain and its graph-based extension, LangGraph. For production, the stack is augmented with observability and evaluation capabilities provided by LangSmith.
Set Up a Smart Wallet
Your agent will require an on-chain identity and a signing capability for transactions. Create an account abstraction using an ERC-4337 wallet or Coinbase’s CDP SDK. Ensure you fund the wallet, engage in DeFi activities, and sign all on-chain activities. Avoid granting an AI agent full access to a treasury wallet. Set a spending limit and a multi-sig security system.
Connect the LLM via API
Integrate your chosen LLM through its API. Set up a system prompt that includes the role, constraints, and output type for your agent. Instruct your model to structure JSON so your middleware can reliably parse actions such as “swap,” “stake,” or “vote.” Ensure to test this layer offline before integrating it with any live contract.
Build the Tool Interface
This is the layer that translates LLM decisions into blockchain calls. Configure the tools that the agent can use, including reading the price of a token from an oracle, making a call on a decentralized exchange router, casting a governance vote, or transferring assets. Each tool should validate its parameters to verify that the agent is not trying to call a contract with invalid calldata.
Keep Record
The default setting of LLMs is that they do not retain memory from one call to the next. This is a difficulty for the trading or portfolio-based agent. Use a vector database or a simple key-value store to provide the agent memory about past decisions, holdings, and strategy parameters. Provide the necessary context at the start of the LLM call.
Implement Guardrails
Transactions on the blockchain are non-reversible. Before you deploy on mainnet, define the maximum transaction size, daily budget, whitelisted contract addresses the agent can communicate with, and a kill switch that a human operator can activate. Establish these thresholds both at the system prompt level of the agent and at the smart contract level. Do not rely solely on the LLM to honor these thresholds.
Test on a Testnet
Using a testnet (such as Base Sepolia or Ethereum Sepolia), execute the full loop of the agent, including reasoning, tool calls, wallet transactions, and outcome logging. Use LangSmith or a similar observability tool to trace every step. Verify that there are no hallucinations during the tool calls, token misuse, or latency problems.
Monitor in Production Environments
Almost 89% of AI agent teams are utilizing observability tooling within their AI agent teams for production environments. This is not optional for on-chain agents since a bad decision results in loss of money. Log all LLM calls, all tool calls, and all transaction hashes. Set up alerts for anomalous behavior such as unusually large swaps or repeated failed transactions.
